Enhanced Iterative Data Refinement for Self-Supervised Seismic Noise Attenuation via Low-Rank Casorati Matrix Representation

Transport in Porous Media · Springer Nature · 2026-09-27 · DOI: 10.1007/s11004-026-10347-0

Аннотация

Random noise attenuation remains a fundamental challenge in seismic data processing, particularly in scenarios where signal-to-noise ratios (SNRs) exhibit significant variability across traces. Conventional low-rank denoising methods often incur prohibitive computational costs due to high-dimensional mappings, while existing self-supervised approaches frequently lack efficient, cost-effective strategies for obtaining accurate noise models without clean data. To address these limitations, this paper proposes an enhanced iterative data refinement (EIDR) framework for self-supervised seismic noise attenuation, which exploits the inherent low-rank property of seismic data via Casorati matrix representation. Theoretically, we rigorously substantiate the noisier–noisy training paradigm within the context of seismic denoising, proving that neural networks trained on appropriately constructed noisier–noisy pairs converge to conditional expectations, thereby yielding statistically optimal denoising estimates under the mean squared error (MSE) criterion. Methodologically, EIDR incorporates three key innovations: (1) a nondestructive dimensionality reduction strategy that rearranges three-dimensional seismic cubes into two-dimensional Casorati matrices while preserving low-rank characteristics, (2) a blind noise estimation approach based on multiple regression theory that generates noise samples with distributions closely matching actual seismic noise, and (3) an ensemble learning mechanism with knowledge distillation that stabilizes training across iterative refinement epochs. Comprehensive experiments on synthetic seismograms demonstrate that EIDR achieves superior performance relative to existing self-supervised methods, achieving peak signal-to-noise ratio (PSNR) improvements of 1.8–6.3 dB, structural similarity index measure (SSIM) gains of 0.04–0.17, and MSE reductions of 17–35% depending on the data domain and noise estimation configuration. Under optimal conditions using Casorati matrix representation with common-offset gather noise estimation, the method achieves relative error reductions exceeding 30%. Field validation further confirmed that EIDR effectively suppresses complex unknown noise while preserving structural fidelity and amplitude integrity.

Читать у издателя →

Похожие статьи